HOL1

advertisement
Theorem Prover HOL, overview
Wishnu Prasetya
wishnu@cs.uu.nl
www.cs.uu.nl/docs/vakken/pv
setting up a proof
applying a proof step
the 2-nd subgoal
the 1-st subgoal
solving the first subgoal
2
Features
 “higher order”, as opposed to a first order prover as
Z3  highly expressive! You can model lots of things
in HOL.
 A huge collection of theories and proof utilities. Well
documented.
 Safe
 computer-checked proofs.
 Simple underlying logic, you can trust it.
 Unless you hack it, you cannot infer falsity.
3
HOL is a DSL embedded in ML
HOL
ML
 ML is a mature functional programming language.
 You can access both HOL and ML
 It gives you powerful meta programming of HOL! E.g. for:
 scripting your proofs
 manipulating your terms
 translating back and forth to other representations
 Interesting options to embed your own DSL/logic
4
Non-features
 The name “theorem prover” is a bit misleading. Higher
order logic is undecidable.
 Don’t expect HOL to do all the work for you!
 It doesn’t have a good incremental learning material.
But once you know your way… it’s really a powerful
thing.
5
ML  programming level
 E.g. these functions in ML:
val zero = 0 ;
fun identity x = x ;
 These are ordinary programs, you can execute them.
E.g. evaluating identity zero will produce 0.
 Unfortunately, there is no direct way to prove
properties of these ML functions. E.g. that identity .
identity = identity
6
HOL level
 We can model them in HOL as follows:
val zero_def
= Define `zero = 0` ;
val identity_def = Define `identity x = x` ;
 The property we want to prove:
--`identity o identity = identity`--
7
The proof (scripted) in HOL
val my_first_theorem = prove (
--`identity o identity = identity`--,
CONV_TAC FUN_EQ_CONV
THEN REWRITE_TAC [o_DEF]
THEN BETA_TAC
THEN REWRITE_TAC [identity_def]
);
But usually you prefer to prove your formula first interactively, and later
on collect your proof steps to make a neat proof script as above.
8
Model and the real thing
 Keep in mind that what we have proven is a theorem
about the models !
 E.g. in the real “identity” in ML :
identity (3/0) = 3/0  exception!
 We didn’t capture this behavior in HOL.
 HOL will say it is true (aside from the fact that x/0 is undefined in
HOL).
 There is no built-in notion of exception in HOL, though we can
model it if we choose so.
9
Core syntax of HOL
 Core syntax is that of simple typed -calculus:
term ::=
|
|
|
var | const
term term
\ var. term
term : type
 Terms are typed.
 We usually interact on its extended syntax, but all
extensions are actually built on this core syntax.
10
Types
 Primitive: bool, num, int, etc.
 Product, e.g.: (1,2) is a term of type num#num
 List, e.g.: [1;2;3] is a term of type num list
 Function, e.g.: (\x. x+1) is a term of type num->num
 Type variables, e.g.:
(\x. x) is a term of type ‘a -> ‘a
 You can define new types.
11
Extended syntax
 Boolean operators:
~p ,
p /\ q ,
p \/ q ,
p ==> q
 Quantifications: // ! = , ? = 
(!x. f x = x) ,
(?x. f x = x)
 Conditional:
if ... then ... else ... // alternatively g -> e1 | e2
 Tuples and lists  you have seen.
 Sets, e.g. {1,2,3}  encoded as int->bool.
12
 You can define your own constants, operators, quantifiers etc.
Examples of modeling in HOL
 Define `double x = x + x` ;
 Define `(sum [ ] = 0)
/\
(sum (x::s) = x + sum s)` ;
 Define `(map f [ ] = [ ])
/\
(map f (x::s) = f x :: map f s)` ;
13
Modeling programs
 Define `skip state = state` ;
 Define `SEQ S1 S2 state = S2 (S1 state)` ;
(higher order at work)
 So, how do we define GCL’s box ?
14
Modeling properties
 How to express that a relation is reflexive and
transitive?
 Define `isReflexive R = (!x. R x x)` ;
(so what is the type of isReflexive?)
 Define `isTransitive R
=
(!x y z. R x y /\ R y z ==> R x y)` ;
 How to define the reflexive and transitive closure of R
15
?
Theorems and proofs
Theorem
 HOL terms: --`0`--
--`x = x`--
 Theorem : a bool-typed HOL term wrapped in a special type called
“thm”, meant to represent a valid fact.
|- x = x
 The type thm is a protected data type, in such a way that you can only
produce an instance of it via a set of ML functions encoding HOL
axioms and primitive inference rules (HOL primitive logic).
 So, if this primitive logic is sound, there is no way a user can produce an
17
invalid theorem.
 This primitive logic is very simple; so you can easily convince yourself of its
soundness.
Theorem in HOL
 More precisely, a theorem is internally a pair (term list
* term), which is pretty printed e.g. like:
[a1, a2, ...]
|-
c
Intended to mean that a1 /\ a2 /\ ...
implies c.
 Terminology: assumptions, conclusion.
 |- c abbreviates [] |- c.
18
Inference rule
 An (inference) rule is essentially a function that produces a theorem.
E.g. this (primitive) inf. rule :
A |- t1  t2 , B |- t1
----------------------------------- Modus Ponens
A @ B |- t2
is implemented by a rule called MP : thmthmthm
 Rules can be composed:
fun myMP t1 t2 = GEN_ALL (MP t1 t2)
19
Backward proving
 Since a “rule” is a function of type (essentially) thmthm,
it implies that to get a theorem you have to “compose” theorems.
 forward proof; you have to work from axioms
 For human it is usually easier to work a proof backwardly.
 HOL has support for backward proving. Concepts :
20
 Goal
 terms representing what you want to prove
 Tactic
 a function that reduce a goal to new goals
Goal
 type goal = term list * term
Pretty printed:
[ a1, a2 , ... ]
?-
h
Represent our intention to prove
[ a1 , a2 , ... ] |- h
 Terminology : assumptions, hypothesis
 type tactic = goal  goal list * proof_func
21
Proof Tree
A proof constructed by applying tactics has in principle a tree structure, where at
every node we also keep the proof function to ‘rebuild’ the node from its children.
g1
If all leaves are ‘closed’ (proven) we build
the root-theorem by applying the proof
functions in the bottom-up way.
tac1
g2
tac2
g3
tac3
(proven)
g4
tac4
(proven)
22
In interactive-proof-mode, such a ‘proof
tree’ is actually implemented as a ‘proof
stack’ (show example).
Proof Function
 type tactic = goal  goal list * proof_func
 So, suppose you have this definition of tac :
tac (A ?- h)
=
( [ A’ ?- h’ ] ,

)
// so, just 1 new subgoal
Then the  has to be such that :

[ A’ |- h’ ]
=
A |- h
 So, a pf is an inference rule, and tac is essentially the reverse of
this rule.
23
Interactive backward proof
(Desc 5.2)
 HOL maintains a global state variable of type proofs :
 proofs
: set of active/unfinished goalstacks
 goalstack : implementation of proof tree as a stack
 A set of basic functions to work on these structures.
 Setting up a new goalstack :
g
: term quotation  proofs
set_goal : goal  proofs
 Applying a tactic to the current goal in the current goalstack:
e (expand) : tactic  goalstack
24
For working on proofs/goalstack...
 Switching focus
r (rotate) : int  goalstack
 Undo
25
b
: unit  goalstack
restart
: unit  goalstack
drop
: unit  proofs
Some basic rules and tactics
Shifting from/to asm... (Old Desc 10.3)
A |- v
------------------------ DISCH u
A / {u} |- u ==> v
A ?- u ==> v
-----------------------DISCH_TAC
A + u ?- v
A |- u ==> v
---------------------- UNDISCH
A + u |- v
A ?- v
------------------------ UNDISCH_TAC u
A / {u} ?- u ==> v
26
Some basic rules and tactics
Modus Ponens (Old Desc 10.3)
A1 |- t
A2 |- t  u
----------------------- MP
A1  A2 |- u
A ?- u
A‘ |- t
-------------------- MP_TAC
A ?- t  u
A1 |- to
A2 |- !x. tx  ux
--------------------------- MATCH_MP
A1  A2 |- uo
A ?- uo
A’ |- !x. tx  ux
--------------------------- MATCH_MP_TAC
A ?- to
A’ should be a subset of A
27
Some basic rules and tactics
Stripping and introducing  (Old Desc 10.3)
A |- !x. P
---------------- SPEC u
A |- P[u/x]
A |- P
---------------- GEN x
A |- !x. P
provided x is not free in A
28
A ?- P
---------------------- SPEC_TAC(u,x)
A ?- !x. P[x/u]
A ?- !x. P x
---------------------- GEN_TAC
A ?- P[x’/x]
x’ is chosen so that it is not bound in A
Some basic rules and tactics
Intro/striping  (Old Desc 10.3)
A |- P
---------------------- EXISTS (?x. P[x/u], u)
A |- ?x. P[x/u]
A ?- ?x. P
--------------------------- EXISTS_TAC u
A ?- P[u/x]
29
Rewriting
(Old Desc 10.3)
A ?- t
------------------- SUBST_TAC [A’ |- u=v]
A ?- t[v/u]
• provides A’  A
• you can supply more equalities...
A ?- t
------------------- REWRITE_TAC [A’ |- u=v]
A ?- t[v/u]
30
• also performs matching e.g. |- f x = x will also match
“... f (x+1)”
• recursive
• may not terminate!
Tactics Combinators (Tacticals)
(Old Desc 10.4)
 The unit and zero 
 ALL_TAC
// a ‘skip’ 
 NO_TAC
// always fail
 Sequencing :
31
 t1 THEN t2
 apply t1, then t2 on all subgoals generated by t1
 t THENL [t1,t2,…]
 apply t, then ti on i-th subgoal generated by t
 REPEAT t
 repeatedly apply t until it fails (!)
Examples
 DISCH_TAC ORELSE GEN_TAC
 REPEAT DISCH_TAC
THEN
THEN
EXISTS_TAC “foo”
ASM_REWRITE_TAC [ ]
 fun UD1 (asms,h)
=
( if null asms then NO_TAC
else UNDISCH_TAC (hd asms) ) (asms,h) ;
32
Practical thing: quoting HOL terms
(Desc 5.1.3)
 Remember that HOL is embedded in ML, so you have to quote HOL
terms; else ML thinks it is a plain ML expression.
 ‘Quotation’ in Moscow ML is essentially just a string:
`x y z`
 is just “x y z”
Notice the backquotes!
 But it is represented a bit differently to support antiquotation:
val aap = 101
` a b c ^aap d e f `
 [QUOTE “a b c", ANTIQUOTE 101, QUOTE “d e f"] : int frag list
33
Quoting HOL terms
 The ML functions Term and Type parse a quotation to ML
“term” and “hol_type”; these are ML datatypes
representing HOL term and HOL type.
Term `identity (x:int)`
 returns a term
Type `:num->num`
 returns a hol_type
Actually, we often just use this alternate notation,
which has the same effect:
--`identity (x:int)`-34
A bit inconsistent styles
 Some functions in HOL expect a term, e.g. :
prove : term -> tactic -> thm
 And some others expect a frag list / quotation 
g : term frag list -> proofs
Define : term frag list -> thm
35
Some common proof techniques
(Desc 5.3 – 5)
 Power tactics
 Proof by case split
 Proof by contradiction
 In-line lemma
 Induction
36
Power Tactics: Simplifier
 Power rewriter, usually to simplify goal :
SIMP_TAC: simpset thm list  tactic
standard simpsets: std_ss, int_ss, list_ss
 Does not fail. May not terminate.
 Being a complex magic box, it is harder to predict what
you get.
 You hope that its behavior is stable over future versions.
37
Examples
 Simplify goal with standard simpset:
SIMP_TAC std_ss [ ]
(what happens if we use list_ss instead?)
 And if you also want to use some definitions to
simplify:
SIMP_TAC std_ss [ foo_def, fi_def , … ]
38
(what’s the type of foo_def ? )
Other variations of SIMP_TAC
 ASM_SIMP_TAC
 FULL_SIMP_TAC
 RW_TAC does a bit more :
 case split on any if-then-else in the hypothesis
 Reducing e.g. (SUC x = SUC y) to (x=y)
 “reduce” let-terms in hypothesis
39
Power Tactics: Automated Provers
 1-st order prover: PROVE_TAC : thm list -> tactic
 Integer arithmetic prover: ARITH_TAC, COOPER_TAC
(from intLib)
 Natural numbers arith. prover:
numLib)
 Undecidable.
 They may fail.
 Magic box.
40
ARITH_CONV
(from
Examples
 Simplify then use automated prover :
RW_TAC std_ss [ foo_def ]
THEN PROVE_TAC [ ]
 In which situations do you want to do these?
RW_TAC std_ss [ foo_def ]
THEN TRY (PROVE_TAC [ ])
RW_TAC std_ss [ foo_def ]
THEN ( PROVE_TAC [ ] ORELSE ARITH_TAC )
41
Case split
 ASM_CASES_TAC : term  tactic
A ?- u
------------------------------ ASM_CASES_TAC t
(1) t + A ?- u
(2) ~t + A ?- u
 Split on data constructors, Cases / Cases_on
A ?- ok s
---------------------------------- Cases_on `s`
(1) A ?- ok [ ]
(2) A ?- ok (x::t)
42
Induction
 Induction over recursive data types: Induct/Induct_on
?- ok s
---------------------------------- Induct_on `s`
(1) ?- ok [ ]
(2) ok t ?- ok (x::t)
 Other types of induction:
 Prove/get the corresponding induction theorem
 Then apply MP
43
Adding “lemma”
 by : (quotation * tactic)  tactic
44
// infix
A ?- t
--------------------------------- lemma by tac
lemma + A ?- t
If tac proves the
lemma
A ?- t
--------------------------------- lemma by tac
(1) lemma + A ?- t
(2) A ?- z
If tac only
reduces lemma
to z
Adding lemma
 But when you use it in an interactive proof perhaps
you want to use it like this:
`foo x > 0` by ALL_TAC
What does this do ?
45
Proof by contradiction
 SPOSE_NOT_THEN : (thmtactic)tactic
SPOSE_NOT_THEN f
 assumes hyp |- hyp.
 now you must prove False.
 f (hyp |- hyp) produces a tactic, this is then applied.
 Example:
A ?- f x = x
----------------------------- SPOSE_NOT_THEN ASSUME_TAC
~(f x = x) + A ?- F
46
Download